David Ondrej open-sources his entire agent-skill library
Hundreds of hours of trial and error, handed over for free. Even if you rewrite half of it, a working reference library beats a blank page every time. Go take what's useful.
This week · July 6โ12, 2026
A weekly handful of the best things I read about AI.
Hundreds of hours of trial and error, handed over for free. Even if you rewrite half of it, a working reference library beats a blank page every time. Go take what's useful.
A cost pattern worth stealing. Let the expensive frontier model do the planning and hand the grunt work to cheaper ones. Obvious in hindsight, and it quietly changes how you budget a project.
The clearest writing yet on why the model stopped being the bottleneck. Thariq's frame: the gap between what you asked for and what the work actually needs is where quality leaks out, and the real skill is surfacing those gaps before, during, and after you build. Worth reading twice.
Hundreds of hours of trial and error, handed over for free. Even if you rewrite half of it, a working reference library beats a blank page every time. Go take what's useful.
A cost pattern worth stealing. Let the expensive frontier model do the planning and hand the grunt work to cheaper ones. Obvious in hindsight, and it quietly changes how you budget a project.
The clearest writing yet on why the model stopped being the bottleneck. Thariq's frame: the gap between what you asked for and what the work actually needs is where quality leaks out, and the real skill is surfacing those gaps before, during, and after you build. Worth reading twice.
shadcn shipping composable components for chat UIs, starting with the streaming conversation layer. Dry on its own, but it matters: as every product bolts on a chat surface, having clean, ownable primitives beats reinventing streaming and message state each time. One for the builders.
A heavy dictation user, top 0.1%, ditched his paid voice tool for an open-source one that runs locally on his Mac with no API key and handles slang better. The broader signal worth watching: local, private, free tools are quietly getting good enough to cancel the subscription. Short and practical.
Anthropic's playbook for working alongside agents as the tools ship into Slack. The interesting shift is treating an agent less like software you run and more like a teammate you delegate to, which changes how you hand off and review work. Read it before your org figures this out the hard way.
A tight framing of the shift everyone's circling: you stop prompting the agent and start building the system that prompts it. Schedule, discover, build, verify, repeat. Worth it as the one-line mental model before you read anything longer on loops.
The whole thing is one habit: every day, Claude builds and prioritizes a to-do list against your actual goals, then calendar-blocks it. Deliberately unimpressive, which is the point. The skills that stick are the boring ones you run every morning, not the clever ones you run once.
The announcement: Claude now joins Slack as a taggable team member for Enterprise and Team plans. Worth noting because Slack, not a terminal, is where most teams already coordinate, and putting the agent there changes who ends up using it. The reference point for the human-agent-teams conversation that followed.
Thariq's field notes on Claude-in-Slack, a genuinely new form factor that nobody has best practices for yet. Early and speculative, but that's the value: you're reading someone figuring out the patterns in real time rather than after they've calcified. Good companion to the launch post.
A pointer to Addy Osmani's walkthrough of how the software lifecycle is changing, phase by phase, from autocomplete to autonomous agents. Osmani is a reliable, non-hypey guide to this stuff, so it's a solid framework read if you want the shape of the shift rather than a hot take.
Boris Cherny on living inside Artifacts: visual explanations of tricky code, system diagrams, dashboards he shares with the team, all built from the session at a private link. The signal here is a power user telling you which everyday habit actually changed how he works. Small feature, big workflow shift.
Andreessen's long, unusually good essay reframing SpaceX as the one company assembling the full stack for a post-scarcity future: cheap launch, orbital compute, lunar industry, Mars. The engineering-culture sections carry it, especially Musk's five-step Algorithm and the idiot index. Not AI-specific, but the orbital-data-center thesis ties directly to where compute and energy are heading. A real sit-down read.
Block's numbers on their internal agent system are the ones to quote in an argument. Engineers tag it in Slack, it researches, plans, and ships: 200,000 operations a day, 1,500 PRs merged a week, 15% of all production code. This is what agents coordinating across a real codebase looks like at scale, not in a demo.
A launch pitch dressed as a provocation: AI is making marketers lazy, so make the website do the work instead. Mostly a product announcement with a $27M seed behind it, but the underlying bet, that the site itself becomes an active agent rather than a brochure, is where a lot of marketing tooling is heading.
A neat idea made real: a feedback loop that catches visual slop and design-system drift while the agent builds, instead of after. It turns design consistency from a thing you ask for into a hook that runs automatically. If your agents produce frontends that look slightly off every time, this is the missing guardrail.
A shared full guide to shipping your first agent, with the pitch that it collapses a two-week slog into a day. Standard on-ramp material, but a decent starting point if you're going zero to one and want a single path instead of ten browser tabs.
Karpathy's line, put to work: stop being the bottleneck, put in few tokens, have a huge amount happen on your behalf. The framing is that loop engineering is the mechanism that actually delivers that. A quick nudge toward the mindset shift, not a how-to.
The best long read on why a top-tier model isn't the point. The system around it is. Loops, memory, verifier sub-agents, and state files are what make each run leave the next one smarter, and the piece lays out the whole stack with cost-routing advice for when to reach for the expensive model versus a cheap one. Long, but it's the map most people are missing.
Steinberger's minimal recipe: an orchestrator that wakes every five minutes and directs work to threads, combined with triage, auto-review, and computer-use skills so work lands on its own. The takeaway is how little scaffolding a useful loop actually needs once you have sharp skills to call. Concrete and copyable.
The demo that makes the capability click: on a customer call, Claude transcribed in the background and built the features the customer was wishing for in real time, ready by the end of the call. One anecdote, but it reframes what fast means. This is the moment a lot of people realized the ground had moved.
A non-engineer spent a billion tokens running Claude's top model on real work across writing, strategy, security, and design. Refreshing because it's a user's honest read on where the model actually delivers, not a benchmark chart. Useful counterweight to the launch-day hype.
An Anthropic engineer on two things the top model changes: self-correction loops and memory. The sharpest takeaway is that the model shouldn't grade its own work. An independent verifier explores harder and recovers from dead ends where self-critique stalls at good enough. Short, concrete, and from someone who actually ran the experiments.
The best explainer on the phrase everyone was repeating without defining. A loop is cron plus a decision-maker in the body: the model, not a hardcoded script, picks the next move each tick. The punchlines land, that the loop, not the model, is now the expensive part, and that the real asset is the skills a loop calls, not the loop itself. If you read one thing on loops, this.
A plain-English tour of the Claude features hiding in plain sight: Projects, memory, extended thinking, scheduled tasks, prompt caching, custom roles. Nothing exotic, but the value is in the framing. Each one takes minutes to set up and pays off daily. Good to forward to anyone still treating Claude as a fancier search box.
Just a pointer to a GitHub repo of agent skills. Low signal on its own, but these skill collections are where the real how-to lives. Clone it, read how the skills are structured, keep what fits.
An honest field report: dynamic workflows felt useless until they were pointed at adversarial review. The move is splitting a review into thin, mean, single-focus lanes, correctness, duplication, safety, each agent required to bring a way to verify its own finding. It beats self-review because the model likes its own work too much. The step-by-step is copy-and-adapt ready.
Anthropic's own writeup on building agents that do data analysis. The useful parts are the unglamorous ones: skills, clean data foundations, and evaluations. If you're wiring an agent into real business numbers, this is the part everyone skips and then regrets.
The line most people skim: every agent in a workflow can run a different model. This is the proof. A 33-agent audit, cheap readers plus one smart synthesizer, ran in under five minutes and found a real buried error. Put the expensive model where the thinking happens and run the rest on the cheap one. That's the economics of parallel agents in one example.
The clearest map of where interfaces are heading: agents drawing the UI in real time instead of describing it. Three patterns, controlled, declarative, open-ended, each with a different failure mode at scale, and most teams pick one by accident. If you build anything agent-facing, this is the decision tree to read before you're locked in.
A clean way to think about agent design: the model is deliberately thin, and intelligence gets pushed outward into memory, skills, and protocols that the harness composes at runtime. The useful question it hands you is where any new capability should live. Good conceptual scaffolding if agents still feel like a bag of tricks.
A short thread on a real problem: when agents do the work, how do you actually understand what got done? The value is in seeing how people inside Anthropic keep oversight without babysitting. Worth a scroll if your agents are starting to outrun your attention.
The canonical piece on dynamic workflows, from the Anthropic engineer who built them. Claude writes its own custom harness on the fly to beat the failure modes of one long context window: laziness, self-preference, goal drift. The example prompts alone are worth the read, and it names the reusable patterns (fan-out, adversarial verify, tournament) you'll see everywhere else. Start here.
A design studio's field report on a real shift: the deliverable is no longer a PDF brand guide but a folder of structured files an agent can build from. The value moves upstream to the thinking, and the sharp bit is what they call magic_trick.md, the one human, left-of-center idea the system can't generate on its own. Best essay here on where human creativity stays scarce.
A sharp look at how youth sports became a $40 billion industry pricing out ordinary families. Not an AI piece, but a clear read on how private capital reshapes a market that used to be a childhood pastime. Worth it if you care about where consolidation goes when it meets kids and weekends.
A pointed demo of where the moat isn't: someone cloned the core of two heavily funded legal-AI products in two weeks and open-sourced it. The argument, that firms should own their application layer instead of renting it forever, is the uncomfortable question every vertical SaaS company is now facing. Provocative, and hard to dismiss.
Karpathy on a workflow more people should steal: point an LLM at a pile of raw sources and let it compile and maintain a markdown wiki you rarely touch by hand. Once it's big enough, you query it like a research assistant, no fancy RAG required. The best part is that your own explorations file back in, so the knowledge base compounds. One of the most-shared AI ideas of the year for a reason.
Karpathy's follow-up carries the sharper point: in the agent era you stop shipping code and start shipping the idea, kept deliberately vague, and the other person's agent builds it for their needs. It's a small reframe with big implications for how software gets distributed. The knowledge-base spec is the worked example.
A genuinely good step-by-step on running Claude Code 24/7 on an always-on Mac Mini so you can text it tasks from anywhere. The best detail is the one nobody mentions: there's no message queue, so a sleeping laptop drops everything, which is the whole case for dedicated hardware. Practical if you want a personal agent that never goes dark.
A crisp reminder of what belongs in your project config. Most hand-written setup was stuff the agent can discover by reading the code. The new approach keeps only what Claude would get wrong without it: security rules, mandatory workflows, non-obvious gotchas. Delete the rest. 458 lines to 68.
The reference for anyone configuring Claude Code seriously. It walks the whole control center: CLAUDE.md, path-scoped rules, and hooks, with the key nuance that instructions are suggestions but hooks are deterministic. Keep CLAUDE.md under 200 lines or adherence drops. This is the one to bookmark.
No reads match. Try another topic or search.